confidence_maps.py

"""Transformers for confidence map generation."""

import tensorflow as tf
import attr
from typing import List, Text, Union, Tuple
from sleap.nn.data.utils import make_grid_vectors
from sleap.nn.data.offset_regression import make_offsets, mask_offsets


def make_confmaps(
    points: tf.Tensor, xv: tf.Tensor, yv: tf.Tensor, sigma: float
) -> tf.Tensor:
    """Make confidence maps from a set of points from a single instance.

    Args:
        points: A tensor of points of shape `(n_nodes, 2)` and dtype `tf.float32` where
            the last axis corresponds to (x, y) pixel coordinates on the image. These
            can contain NaNs to indicate missing points.
        xv: Sampling grid vector for x-coordinates of shape `(grid_width,)` and dtype
            `tf.float32`. This can be generated by
            `sleap.nn.data.utils.make_grid_vectors`.
        yv: Sampling grid vector for y-coordinates of shape `(grid_height,)` and dtype
            `tf.float32`. This can be generated by
            `sleap.nn.data.utils.make_grid_vectors`.
        sigma: Standard deviation of the 2D Gaussian distribution sampled to generate
            confidence maps.

    Returns:
        Confidence maps as a tensor of shape `(grid_height, grid_width, n_nodes)` of
        dtype `tf.float32`.

        Each channel of the confidence maps will contain the unnormalized PDF of a 2D
        Gaussian distribution with a mean centered at the coordinates of the
        corresponding point, and diagonal covariance matrix (i.e., the same standard
        deviation for both dimensions).

        When the point is perfectly aligned to the sampling grid, the value at that grid
        coordinate is 1.0 since the PDF is not normalized.

        If a point was missing (indicated by NaNs), the corresponding channel will
        contain all zeros.

    See also: sleap.nn.data.make_grid_vectors, make_multi_confmaps
    """
    x = tf.reshape(tf.gather(points, [0], axis=1), [1, 1, -1])
    y = tf.reshape(tf.gather(points, [1], axis=1), [1, 1, -1])
    cm = tf.exp(
        -((tf.reshape(xv, [1, -1, 1]) - x) ** 2 + (tf.reshape(yv, [-1, 1, 1]) - y) ** 2)
        / (2 * sigma ** 2)
    )
    cm = tf.math.maximum(0.0, cm)  # Replaces NaNs with 0.
    return cm


def make_multi_confmaps(
    instances: tf.Tensor, xv: tf.Tensor, yv: tf.Tensor, sigma: float
) -> tf.Tensor:
    """Make confidence maps for multiple instances through reduction.

    Args:
        instances: A tensor of shape `(n_instances, n_nodes, 2)` and dtype `tf.float32`
            containing instance points where the last axis corresponds to (x, y) pixel
            coordinates on the image. This must be rank-3 even if a single instance is
            present.
        xv: Sampling grid vector for x-coordinates of shape `(grid_width,)` and dtype
            `tf.float32`. This can be generated by
            `sleap.nn.data.utils.make_grid_vectors`.
        yv: Sampling grid vector for y-coordinates of shape `(grid_height,)` and dtype
            `tf.float32`. This can be generated by
            `sleap.nn.data.utils.make_grid_vectors`.
        sigma: Standard deviation of the 2D Gaussian distribution sampled to generate
            confidence maps.

    Returns:
        Confidence maps as a tensor of shape `(grid_height, grid_width, n_nodes)` of
        dtype `tf.float32`.

        Each channel will contain the elementwise maximum of the confidence maps
        generated from all individual points for the associated node.

    Notes:
        The confidence maps are computed individually for each instance and immediately
        max-reduced to avoid maintaining the entire set of all instance maps. This
        enables memory-efficient generation of multi-instance maps for examples with
        large numbers of instances.

    See also: sleap.nn.data.make_grid_vectors, make_confmaps
    """
    # Initialize output tensors.
    grid_height = tf.shape(yv)[0]
    grid_width = tf.shape(xv)[0]
    n_nodes = tf.shape(instances)[1]
    cms = tf.zeros((grid_height, grid_width, n_nodes), tf.float32)

    # Eliminate instances completely outside of image.
    in_img = (instances > 0) & (
        instances < tf.reshape(tf.stack([xv[-1], yv[-1]], axis=0), [1, 1, 2])
    )
    in_img = tf.reduce_any(tf.reduce_all(in_img, axis=-1), axis=1)
    in_img = tf.ensure_shape(in_img, [None])
    instances = tf.boolean_mask(instances, in_img)

    # Generate and reduce outputs by instance.
    for points in instances:
        cms_instance = make_confmaps(points, xv, yv, sigma=sigma)
        cms = tf.maximum(cms, cms_instance)

    return cms


def make_multi_confmaps_with_offsets(
    instances: tf.Tensor,
    xv: tf.Tensor,
    yv: tf.Tensor,
    stride: int,
    sigma: float,
    offsets_threshold: float,
) -> tf.Tensor:
    """Make confidence maps and offsets for multiple instances through reduction.

    Args:
        instances: A tensor of shape `(n_instances, n_nodes, 2)` and dtype `tf.float32`
            containing instance points where the last axis corresponds to (x, y) pixel
            coordinates on the image. This must be rank-3 even if a single instance is
            present.
        xv: Sampling grid vector for x-coordinates of shape `(grid_width,)` and dtype
            `tf.float32`. This can be generated by
            `sleap.nn.data.utils.make_grid_vectors`.
        yv: Sampling grid vector for y-coordinates of shape `(grid_height,)` and dtype
            `tf.float32`. This can be generated by
            `sleap.nn.data.utils.make_grid_vectors`.
        stride: Scaling factor for offset coordinates. The individual offset vectors
            will be divided by this value. Useful for adjusting for strided sampling
            grids so that the offsets point to the smaller grid coordinates.
        sigma: Standard deviation of the 2D Gaussian distribution sampled to generate
            confidence maps.
        offsets_threshold: Minimum confidence map value below which offsets will be
            replaced with zeros.
        flatten_offsets: If `True`, the last two channels of the offset maps will be
            flattened to produce rank-3 tensors. If `False`, the generated offset maps
            will be rank-4 with shape `(height, width, n_nodes, 2)`.

    Returns:
        A tuple of `(confmaps, offsets)`.

        `confmaps` are confidence maps as a tensor of shape
        `(grid_height, grid_width, n_nodes)` and dtype `tf.float32`.

        Each channel will contain the elementwise maximum of the confidence maps
        generated from all individual points for the associated node.

        `offsets` are offset maps as a `tf.Tensor` of shape
        `(grid_height, grid_width, n_nodes, 2)` and dtype `tf.float32`. The last axis
        corresponds to the x- and y-offsets at each grid point for each node.

    Notes:
        The confidence maps and offsets are computed individually for each instance
        and immediately max-reduced to avoid maintaining the entire set of all instance
        maps. This enables memory-efficient generation of multi-instance maps for
        examples with large numbers of instances.

    See also: sleap.nn.data.make_grid_vectors, make_confmaps, make_multi_confmaps
    """
    # Initialize output tensors.
    grid_height = tf.shape(yv)[0]
    grid_width = tf.shape(xv)[0]
    n_nodes = tf.shape(instances)[1]
    cms = tf.zeros((grid_height, grid_width, n_nodes), tf.float32)
    offsets = tf.zeros((grid_height, grid_width, n_nodes, 2), tf.float32)

    # Eliminate instances completely outside of image.
    in_img = (instances > 0) & (
        instances < tf.reshape(tf.stack([xv[-1], yv[-1]], axis=0), [1, 1, 2])
    )
    in_img = tf.reduce_any(tf.reduce_all(in_img, axis=-1), axis=1)
    in_img = tf.ensure_shape(in_img, [None])
    instances = tf.boolean_mask(instances, in_img)

    # Generate and reduce outputs by instance.
    for points in instances:
        cms_instance = make_confmaps(points, xv, yv, sigma=sigma)
        cms = tf.maximum(cms, cms_instance)
        offsets_instance = mask_offsets(
            make_offsets(points, xv, yv, stride=stride),
            cms_instance,
            threshold=offsets_threshold,
        )
        offsets += offsets_instance

    return cms, offsets


@attr.s(auto_attribs=True)
class MultiConfidenceMapGenerator:
    """Transformer to generate multi-instance confidence maps.

    Attributes:
        sigma: Standard deviation of the 2D Gaussian distribution sampled to generate
            confidence maps. This defines the spread in units of the input image's grid,
            i.e., it does not take scaling in previous steps into account.
        output_stride: Relative stride of the generated confidence maps. This is
            effectively the reciprocal of the output scale, i.e., increase this to
            generate confidence maps that are smaller than the input images.
        centroids: If `True`, generate confidence maps for centroids rather than
            instance points.
        with_offsets: If `True`, also return offsets for refining the peaks.
        offsets_threshold: Minimum confidence map value below which offsets will be
            replaced with zeros.
        flatten_offsets: If `True`, the last two channels of the offset maps will be
            flattened to produce rank-3 tensors. If `False`, the generated offset maps
            will be rank-4 with shape `(height, width, n_nodes, 2)`.
    """

    sigma: float = 1.0
    output_stride: int = 1
    centroids: bool = False
    with_offsets: bool = False
    offsets_threshold: float = 0.2
    flatten_offsets: bool = True

    @property
    def input_keys(self) -> List[Text]:
        """Return the keys that incoming elements are expected to have."""
        if self.centroids:
            return ["image", "centroids"]
        else:
            return ["image", "instances"]

    @property
    def output_keys(self) -> List[Text]:
        """Return the keys that outgoing elements will have."""
        if self.centroids:
            keys = self.input_keys + ["centroid_confidence_maps"]
        else:
            keys = self.input_keys + ["confidence_maps"]
        if self.with_offsets:
            keys.append("offsets")

        return keys

    def transform_dataset(self, input_ds: tf.data.Dataset) -> tf.data.Dataset:
        """Create a dataset that contains the generated confidence maps.

        Args:
            input_ds: A dataset with elements that contain the keys `"image"`, `"scale"`
                and either "instances" or "centroids" depending on whether the
                `centroids` attribute is set to `True`.

        Returns:
            A `tf.data.Dataset` with the same keys as the input, as well as a
            `"confidence_maps"` or `"centroid_confidence_maps"` key containing the
            generated confidence maps.

            If the `with_offsets` attribute is `True`, example will contain a
            `"offsets"` key.

        Notes:
            The output stride is relative to the current scale of the image. To map
            points on the confidence maps to the raw image, first multiply them by the
            output stride, and then scale the x- and y-coordinates by the `"scale"` key.

            Importantly, the `sigma` will be proportional to the current image grid, not
            the original grid prior to scaling operations.
        """
        # Infer image dimensions to generate the full scale sampling grid.
        test_example = next(iter(input_ds))
        image_height = test_example["image"].shape[0]
        image_width = test_example["image"].shape[1]

        # Generate sampling grid vectors.
        xv, yv = make_grid_vectors(
            image_height=image_height,
            image_width=image_width,
            output_stride=self.output_stride,
        )

        def generate_multi_confmaps(example):
            """Local processing function for dataset mapping."""
            if self.centroids:
                points = tf.expand_dims(example["centroids"], axis=1)
                cm_key = "centroid_confidence_maps"
            else:
                points = example["instances"]
                cm_key = "confidence_maps"

            if self.with_offsets:
                cms, offsets = make_multi_confmaps_with_offsets(
                    points,
                    xv,
                    yv,
                    self.output_stride,
                    self.sigma * self.output_stride,
                    self.offsets_threshold,
                )
                example["offsets"] = offsets
                if self.flatten_offsets:
                    shape = tf.shape(example["offsets"])
                    example["offsets"] = tf.reshape(
                        example["offsets"], [shape[0], shape[1], -1]
                    )
            else:
                cms = make_multi_confmaps(
                    points, xv, yv, self.sigma * self.output_stride
                )

            example[cm_key] = cms
            return example

        # Map transformation.
        output_ds = input_ds.map(
            generate_multi_confmaps, num_parallel_calls=tf.data.experimental.AUTOTUNE
        )
        return output_ds


@attr.s(auto_attribs=True)
class InstanceConfidenceMapGenerator:
    """Transformer to generate instance-centered confidence maps.

    Attributes:
        sigma: Standard deviation of the 2D Gaussian distribution sampled to generate
            confidence maps. This defines the spread in units of the input image's grid,
            i.e., it does not take scaling in previous steps into account.
        output_stride: Relative stride of the generated confidence maps. This is
            effectively the reciprocal of the output scale, i.e., increase this to
            generate confidence maps that are smaller than the input images.
        all_instances: If `True`, will also generate the multi-instance confidence maps.
        with_offsets: If `True`, also return offsets for refining the peaks.
        offsets_threshold: Minimum confidence map value below which offsets will be
            replaced with zeros.
        flatten_offsets: If `True`, the last two channels of the offset maps will be
            flattened to produce rank-3 tensors. If `False`, the generated offset maps
            will be rank-4 with shape `(height, width, n_nodes, 2)`.
    """

    sigma: float = 1.0
    output_stride: int = 1
    all_instances: bool = False
    with_offsets: bool = False
    offsets_threshold: float = 0.2
    flatten_offsets: bool = True

    @property
    def input_keys(self) -> List[Text]:
        """Return the keys that incoming elements are expected to have."""
        if self.all_instances:
            return ["instance_image", "center_instance", "all_instances"]
        else:
            return ["instance_image", "center_instance"]

    @property
    def output_keys(self) -> List[Text]:
        """Return the keys that outgoing elements will have."""
        keys = self.input_keys + ["instance_confidence_maps"]
        if self.with_offsets:
            keys.append("offsets")
        if self.all_instances:
            keys.append("all_instance_confidence_maps")
            if self.with_offsets:
                keys.append("all_instance_offsets")
        return keys

    def transform_dataset(self, input_ds: tf.data.Dataset) -> tf.data.Dataset:
        """Create a dataset that contains the generated confidence maps.

        Args:
            input_ds: A dataset with elements that contain the keys `"instance_image"`,
                `"center_instance"` and, if the attribute `all_instances` is `True`,
                `"all_instances"`.

        Returns:
            A `tf.data.Dataset` with the same keys as the input, as well as
            `"instance_confidence_maps"` and, if the attribute `all_instances` is
            `True`, `"all_instance_confidence_maps"` keys containing the generated
            confidence maps.

            If the `with_offsets` attribute is `True`, example will contain a
            `"offsets"` key.

        Notes:
            The output stride is relative to the current scale of the image. To map
            points on the confidence maps to the raw image, first multiply them by the
            output stride, and then scale the x- and y-coordinates by the `"scale"` key.

            Importantly, the `sigma` will be proportional to the current image grid, not
            the original grid prior to scaling operations.
        """
        # Infer image dimensions to generate sampling grid.
        test_example = next(iter(input_ds))
        image_height = test_example["instance_image"].shape[0]
        image_width = test_example["instance_image"].shape[1]

        # Generate sampling grid vectors.
        xv, yv = make_grid_vectors(
            image_height=image_height,
            image_width=image_width,
            output_stride=self.output_stride,
        )

        def generate_confmaps(example):
            """Local processing function for dataset mapping."""
            example["instance_confidence_maps"] = make_confmaps(
                example["center_instance"],
                xv=xv,
                yv=yv,
                sigma=self.sigma * self.output_stride,
            )

            if self.with_offsets:
                example["offsets"] = mask_offsets(
                    make_offsets(
                        example["center_instance"], xv, yv, stride=self.output_stride
                    ),
                    example["instance_confidence_maps"],
                    self.offsets_threshold,
                )
                if self.flatten_offsets:
                    shape = tf.shape(example["offsets"])
                    example["offsets"] = tf.reshape(
                        example["offsets"], [shape[0], shape[1], -1]
                    )

            if self.all_instances:
                if self.with_offsets:
                    cms, offsets = make_multi_confmaps_with_offsets(
                        example["all_instances"],
                        xv,
                        yv,
                        self.output_stride,
                        self.sigma * self.output_stride,
                        offsets_threshold=self.offsets_threshold,
                    )
                    example["all_instance_offsets"] = offsets
                    if self.flatten_offsets:
                        shape = tf.shape(example["all_instance_offsets"])
                        example["all_instance_offsets"] = tf.reshape(
                            example["all_instance_offsets"], [shape[0], shape[1], -1]
                        )
                else:
                    cms = make_multi_confmaps(
                        example["all_instances"],
                        xv=xv,
                        yv=yv,
                        sigma=self.sigma * self.output_stride,
                    )
                example["all_instance_confidence_maps"] = cms

            return example

        # Map transformation.
        output_ds = input_ds.map(
            generate_confmaps, num_parallel_calls=tf.data.experimental.AUTOTUNE
        )
        return output_ds


@attr.s(auto_attribs=True)
class SingleInstanceConfidenceMapGenerator:
    """Transformer to generate single-instance confidence maps.

    Attributes:
        sigma: Standard deviation of the 2D Gaussian distribution sampled to generate
            confidence maps. This defines the spread in units of the input image's grid,
            i.e., it does not take scaling in previous steps into account.
        output_stride: Relative stride of the generated confidence maps. This is
            effectively the reciprocal of the output scale, i.e., increase this to
            generate confidence maps that are smaller than the input images.
        with_offsets: If `True`, also return offsets for refining the peaks.
        offsets_threshold: Minimum confidence map value below which offsets will be
            replaced with zeros.
        flatten_offsets: If `True`, the last two channels of the offset maps will be
            flattened to produce rank-3 tensors. If `False`, the generated offset maps
            will be rank-4 with shape `(height, width, n_nodes, 2)`.
    """

    sigma: float = 1.0
    output_stride: int = 1
    with_offsets: bool = False
    offsets_threshold: float = 0.2
    flatten_offsets: bool = True

    @property
    def input_keys(self) -> List[Text]:
        """Return the keys that incoming elements are expected to have."""
        return ["image", "instances"]

    @property
    def output_keys(self) -> List[Text]:
        """Return the keys that outgoing elements will have."""
        keys = self.input_keys + ["points", "confidence_maps"]
        if self.with_offsets:
            keys.append("offsets")
        return keys

    def transform_dataset(self, input_ds: tf.data.Dataset) -> tf.data.Dataset:
        """Create a dataset that contains the generated confidence maps.

        Args:
            input_ds: A dataset with elements that contain the keys `"instances"` and
                `"image"`.

        Returns:
            A `tf.data.Dataset` with the same keys as the input, as well as
            `"confidence_maps"` containing the generated confidence maps.

        Notes:
            The output stride is relative to the current scale of the image. To map
            points on the confidence maps to the raw image, first multiply them by the
            output stride, and then scale the x- and y-coordinates by the "scale" key.

            Importantly, the `sigma` will be proportional to the current image grid, not
            the original grid prior to scaling operations.
        """
        # Infer image dimensions to generate sampling grid.
        test_example = next(iter(input_ds))
        image_height = test_example["image"].shape[0]
        image_width = test_example["image"].shape[1]

        # Generate sampling grid vectors.
        xv, yv = make_grid_vectors(
            image_height=image_height,
            image_width=image_width,
            output_stride=self.output_stride,
        )

        def generate_confmaps(example):
            """Local processing function for dataset mapping."""
            # Pull out first instance as (n_nodes, 2) tensor.
            example["points"] = tf.gather(example["instances"], 0, axis=0)

            # Generate confidence maps.
            example["confidence_maps"] = make_confmaps(
                example["points"], xv=xv, yv=yv, sigma=self.sigma * self.output_stride
            )

            if self.with_offsets:
                example["offsets"] = mask_offsets(
                    make_offsets(example["points"], xv, yv, stride=self.output_stride),
                    example["confidence_maps"],
                    self.offsets_threshold,
                )
                if self.flatten_offsets:
                    shape = tf.shape(example["offsets"])
                    example["offsets"] = tf.reshape(
                        example["offsets"], [shape[0], shape[1], -1]
                    )

            return example

        # Map transformation.
        output_ds = input_ds.map(
            generate_confmaps, num_parallel_calls=tf.data.experimental.AUTOTUNE
        )
        return output_ds